Animatable Implicit Neural Representations for Creating Realistic Avatars From Videos.

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Reconstructing animatable human models from multi-view video is challenging.
• Existing methods using canonical neural radiance fields and deformation fields are under-constrained and lack motion control.

Purpose of the Study:

• To develop a novel method for reconstructing animatable human models from multi-view video.
• To enable explicit control over human model animation using skeletal motions.

Main Methods:

• Introduced blend weight fields based on skeleton-driven deformation for generating correspondences.
• Represented human geometry as a signed distance field in canonical space.
• Utilized a neural point displacement field to enhance motion modeling.

Main Results:

• The proposed blend weight fields regularize deformation field learning using observable 3D human skeletons.
• The method allows for animating the human model by combining blend weight fields with input skeletal motions.
• Significantly outperforms recent human modeling methods in experimental evaluations.

Conclusions:

• The novel approach effectively reconstructs animatable human models from multi-view video.
• Blend weight fields offer a controllable and regularized alternative for human motion modeling.
• This method advances the state-of-the-art in dynamic human modeling and animation.